Writing in New York Times Magazine, Siddhartha Mukherjee, a physician and author, explores the difficulty of predicting patient death by recounting the "death-sniffing skills" of a nursing home cat—and the potential of a new algorithm designed to help doctors identify the palliative care "sweet spot."
Death: 'The ultimate black box'
Doctors, Mukherjee writes, "have an abysmal track record of predicting which of our patients are going to die." He cites a survey from researchers at University College London that found while "some doctors predicted deaths accurately," other physicians "underestimated death by nearly three months; yet others overestimated it by an equal magnitude."
In fact, Mukherjee spotlights the story of a black-and-white cat, named Oscar, who "was apparently better than most doctors at predicting when a terminally ill patient was about to die." A case study of the cat appeared in the New England Journal of Medicine in 2007, and explains how Oscar would sniff the air and curl up around patients near death at the Steere House nursing home in Rhode Island. Oscar correctly identified 50 near-death patients, Mukherjee writes, but "no one knows how the cat acquired his formidable death-sniffing skills."
Touching on his own experience incorrectly assessing the likelihood of death of one of his patients, Mukherjee states, "Death is our ultimate black box."
An algorithm that predicts death
"But what if an algorithm could predict death?" Mukherjee posits, spotlighting a project spearheaded by Anand Avati, a computer science graduate student at Stanford University, along with a team from the university's medical school.
Avati and his team set out to "teach" an algorithm how to identify which patients were going to die within three to 12 months, as that timeframe is optimal for providing palliative care. Such information "would allow doctors to use medical interventions more appropriately and more humanely," Mukherjee explains. Plus, with the algorithm, "palliative-care teams would be relieved from having to manually scour charts, hunting for those most likely to benefit."
To teach the algorithm, Avati and his team relied on hospital medical record data from a sample of about 200,000 patients with illnesses, such as heart disease and cancer, and used the information as a "proxy time machine." For instance, the team would identify when a patient passed away and then would scroll back in his or her medical records to "collect and analyze medical information before" the optimal palliative care timeframe for that patient in order to find what information "would enable a doctor to predict a demise in that three-to-12-month section of time." The team fed the data—which was both "objective" and "standardized across patients"—into a kind of software architecture called a "deep neural network," which then "adjust[ed] the weights and strengths of each piece of information ... to generate a probability score that a given patient would die within three to 12 months," Mukherjee explains.
After "absorb[ing] information from the nearly 160,000 patients to train itself," the algorithm was used on the remaining 40,000 patients—with surprisingly solid results, Mukherjee writes. Overall, "the false-alarm rate was low," Mukherjee states, noting that about 90% of patients whom the algorithm predicted would die within three to 12 months did, and 95% given a low probability of death survived for longer than 12 months. However, "the rub" of the algorithm is that "it cannot tell us why it has learned," Mukherjee says—the question of "why" remains, "like death, another black box."
Nonetheless, the box can shed light on "expected and unexpected patterns," Mukherjee writes. For instance, Mukherjee explains one patient who died within the predicted window had more obvious signs of impending death, such as bladder and prostate cancer as well as 60 days of hospitalization. But at the same time, the algorithm gave "a surprising amount of weight" to "the fact that scans were made of his spine and that a catheter had been used in his spinal cord," which Mukherjee says are "features that [he] and [his] colleagues might not have recognized as predictors of dying."
'Inherent discomfort' lingers
Despite the promise of such an algorithm, however, Mukherjee says he "cannot shake some inherent discomfort with the thought that an algorithm might understand patterns of mortality better than most humans."
And why, he concludes, "would such a program seem so much more acceptable if it had come wrapped in a black-and-white fur box that, rather than emitting probabilistic outputs, curled up next to us with retracted claws?" (Mukherjee, New York Times Magazine, 1/3).
5 characteristics of programs that capture the full value of palliative care
What many don't realize is the best palliative care programs also yield longer survival rates, lower costs, and improved quality.
In profiling palliative care initiatives around the country, we identified five characteristics of the most effective programs.